AI Coding Weekly: Claude Code, Devin Desktop & More

The AI coding tool ecosystem moves fast. This week brought a flurry of meaningful updates—from Claude Code tightening sandbox security to OpenCode crossing a major adoption milestone and Windsurf completing a full rebranding. Here's everything worth knowing as of June 24, 2026.

Claude Code v2.1.186–187: Sandbox Security & MCP CLI Auth

Anthropic shipped two back-to-back Claude Code releases last week. Version 2.1.187 (June 23) introduced a new sandbox.credentials setting that prevents sandboxed shell commands from reading credential files and secret environment variables—a meaningful security boundary for teams running untrusted or AI-generated scripts. The same release also fixed remote MCP tool calls that previously hung for up to five minutes before timing out, now aborted cleanly with a descriptive error. Org-configured model restrictions can now limit which models appear in the picker, --model flag, and ANTHROPIC_MODEL env var—simplifying governance at scale. See the full Claude Code changelog for the complete list.

Version 2.1.186 (June 22) added two new CLI commands—claude mcp login <name> and claude mcp logout <name>—making it possible to authenticate MCP servers without navigating the interactive menu. Background subagents now surface permission prompts in the main session instead of silently auto-denying them, which removes a frustrating failure mode in complex multi-agent workflows. The bash ! shortcut now automatically triggers a Claude response, and the /review <pr> command was upgraded to use the same engine as /code-review medium.

Auto Mode Now Blocks Destructive Git Commands

Version 2.1.183 (June 19) introduced safety guardrails many agentic developers had been requesting. Claude Code's auto mode now blocks commands that discard work unless the user explicitly asked for it—covering git reset --hard, git checkout -- ., git clean -fd, git stash drop, and git commit --amend when the commit wasn't made by the agent this session. Infrastructure destruction commands (terraform destroy, pulumi destroy, cdk destroy) are similarly blocked unless a specific stack was named. These guardrails significantly reduce the risk of autonomous agents silently deleting uncommitted work.

OpenCode Reaches 160,000 GitHub Stars

The open-source, model-agnostic coding agent OpenCode crossed 160,000 GitHub stars and 7.5 million monthly active developers in June 2026—a remarkable milestone for a tool that spans terminal, desktop, and IDE surfaces. Version 1.17.9 (June 21) improved reliability by forcing a final text response when configured step limits are hit instead of failing mid-run. Session timelines load significantly faster with less flicker, and OpenAI-compatible providers now accept MCP tool schemas that previously failed validation.

A key differentiator is OpenCode Zen—a curated, benchmarked model list with transparent pay-as-you-go pricing ($20 balance, auto top-up, zero markup on inference). It's provider-agnostic and usable with any agent, making it a compelling alternative to opaque platform subscriptions for developers who want cost clarity.

Windsurf Rebrands as Devin Desktop

On June 2, Cognition retired the Windsurf brand and relaunched the IDE as Devin Desktop, with the Agent Command Center as the default surface. The rebrand accompanies support for the open Agent Client Protocol (ACP), positioning Devin Desktop as an orchestration hub rather than a pure IDE. For existing Windsurf users the transition is seamless—same product, new identity, and an expanded agentic mission built around the Devin brand that Cognition has been cultivating.

Cursor Restructures Teams Pricing

Cursor—now exceeding $500M in annualised revenue—restructured its Teams tier into two seat types: Standard ($32/seat/month annual, $40/month monthly) and Premium ($96/seat/month annual, $120/month monthly) with 5× the Standard usage allowance. The split gives engineering managers clearer cost controls—power users can opt into higher limits without forcing the entire team onto the expensive tier. This change comes alongside GitHub Copilot's new $100/month Max plan, signalling an industry-wide move toward tiered, usage-based pricing in AI coding tools.

GitHub Copilot Agent Mode Goes Mainstream

GitHub Copilot's agentic coding mode now works directly through GitHub Issues. Assign an issue to Copilot and it autonomously explores the repository in a GitHub Actions environment, writes code, runs tests, and opens a pull request. Combined with the new $100/month Max plan and flex billing, this positions Copilot as a serious competitor for teams already embedded in the GitHub ecosystem—going head-to-head with Claude Code and OpenCode on fully autonomous coding workflows.

Anthropic Project Glasswing: AI Security at Scale

Beyond the IDE space, Anthropic's Project Glasswing deserves attention. Using the Claude Mythos Preview model, the initiative scanned over 1,000 open-source projects in its first month, surfacing more than 10,000 high- or critical-severity vulnerabilities—plus an estimated 6,202 additional bugs across the broader dataset. This signals a shift: AI coding tools can contribute to the security health of the wider software ecosystem, not just individual developer productivity.

Resources & Further Reading

Claude Code Official Changelog — full release notes for v2.1.181 through v2.1.187.

Claude Code GitHub CHANGELOG.md — community-facing version history on GitHub.

OpenCode on GitHub — open-source codebase, releases, and community discussions.

2026 AI Coding Agent Comparison: Claude Code vs Cursor vs Devin vs Copilot vs Kiro — pricing, features, and agentic capability breakdown.

Best AI Coding Agents Scored Leaderboard (MorphLLM) — live benchmark rankings updated regularly.

OpenCode 160K Stars Deep Dive — analysis of OpenCode's growth trajectory and community adoption.